Agents

An agent file should be more than a boundary list.

Signalane treats agent instructions as collaboration architecture: anchor, role, current truth, evidence, handoff, and recovery path.

Mini thesis

Agents do not stay reliable because a file says "do this" and "do not do that."

Boundary lists matter. They stop obvious mistakes. But serious agent work needs more than permissions and prohibitions, because the hardest failures are often not dramatic violations. They are quiet drifts.

An agent can follow an old scope after the human decision has moved. It can preserve a handoff while losing the reason behind it. It can obey the file and still misunderstand the work. That is why Signalane treats agent files as collaboration constitutions, not static command sheets.

A useful agent file names the human anchor, the working role, the current-truth rule, the evidence standard, the communication rhythm, and the recovery path. It tells the agent how to return to the human when the system becomes unclear, instead of pretending the old document still knows what is true.

The goal is not to make agents more theatrical or more autonomous on paper. The goal is to make them more accountable, more readable, and more able to cooperate without losing the person who gives the work meaning.

What changes

The agent learns to check the work before serving the artifact.

In Signalane, the point of an agent file is not to hypnotize the agent into obedience. It is to give the agent enough anchored context to read a request, compare it with current truth, notice when something is stale, and return to the human before damage is done.